Morphology-Aware Span-First Structured ABSA for Uzbek: Boundary Modeling and a Protocol for Cross-Lingual Turkic Transfer
Year: 2026
Type: preprint
Abstract: Recovering aspect mentions is insufficient for structured sentiment analysis when their opinions, occurrences, or attributes are wrong. We study these errors in Uzbek educational texts using UzEDU-ABSA: 7,440 records, 5,683 texts, and 5,622 allocation groups. MorphoSpan-Uz augments XLM-R with a token-level byte convolution, eight transparent suffix features, and auxiliary endpoint supervision with span–endpoint consistency. It maps contiguous word spans to original Unicode offsets and predicts six-field tuples. We reassess eight archived systems across five seeds and independently reproduce their saved final-test metrics. On 744 records from 567 texts and 564 groups, mean candidate aspect exact F1 is 73.11 ± 14.12 and strict tuple F1 is 26.15 ± 8.16, versus 65.61 ± 18.79 and 19.09 ± 7.44 for XLM-R span. Scores are percentages and dispersions are sample standard deviations across seeds. Retrospective paired hierarchical bootstrap intervals for the improvements are [2.14, 14.49] and [4.35, 9.93] percentage points. Byte features and boundary supervision improve observed means; morphology gains are smaller and metric-dependent. Target-only low-rank adaptation (LoRA) produces no candidates, and constrained generators remain syntactically valid despite low tuple accuracy. Offset annotations have mixed review provenance: 525 records have documented human span confirmation. Conclusions are therefore conditional on annotation quality; prediction-level reproducibility does not establish checkpoint reproducibility. Turkish-to-Uzbek transfer is specified as a prospective protocol; no cross-lingual gain is claimed from the completed experiments.
Source: Research Square
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225787392